Atypical Prosodic Structure as an Indicator of Reading Level and Text Difficulty

Proceedings of NAACL-HLT 2013, pages 715–720, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Atypical Prosodic Structure as an Indicator of Reading Level and Text Difficulty Julie Medero and Mari Ostendorf Electrical Engineering Department University of Washington Seattle, WA 98195 USA {jmedero,ostendor}@uw.edu Abstract Automatic assessment of reading ability builds on applying speech recognition tools to oral reading, measuring words correct per minute. This work looks at more fine-grained analysis that accounts for effects of prosodic context using a large corpus of read speech from a literacy study. Experiments show that lower-level readers tend to produce relatively more lengthening on words that are not likely to be final in a prosodic phrase, i.e. in less appropriate locations. The results have impli- cations for automatic assessment of text dif- ficulty in that locations of atypical prosodic lengthening are indicative of difficult lexical items and syntactic constructions. 1 Introduction Fluent reading is known to be a good indicator of reading comprehension, especially for early readers (Rasinski, 2006), so oral reading is often used to evaluate a student’s reading level. One method that can be automated with speech recognition technol- ogy is the number of words that a student can read correctly of a normed passage, or Words Correct Per Minute (WCPM) (Downey et al., 2011). Since WCPM depends on speaking rate as well as liter- acy, we are interested in identifying new measures that can be automatically computed for use in com- bination with WCPM to provide a better assessment of reading level. In particular, we investigate fine- grained measures that, if useful in identifying points of difficulty for readers, can lead to new approaches for assessing text difficulty. The WCPM is reduced when a person repeats or incorrectly reads a word, but also when they in- troduce pauses and articulate words more slowly. Pauses and lengthened articulation can be an indi- cator of uncertainty for a low-level reader, but these phenomena are also used by skilled readers to mark prosodic phrase structure, facilitating comprehen- sion in listeners. Since prosodic phrase boundaries tend to occur in locations that coincide with certain syntactic constituent boundaries, it is possible to au- tomatically predict prosodic phrase boundary loca- tions from part-of-speech labels and syntactic struc- ture with fairly high reliability for read news stories (Ananthakrishnan and Narayanan, 2008). Thus, we hypothesize that we can more effectively leverage word-level articulation and pause information by fo- cusing on words that are less likely to be associ- ated with prosodic phrase boundaries. By compar- ing average statistics of articulation rate and paus- ing for words at boundary vs. non-boundary loca- tions, we hope to obtain a measure that could aug- ment reading rate for evaluating reading ability. We also hypothesize that the specific locations of hesita- tion phenomena (word lengthening and pausing) ob- served for multiple readers will be indicative of par- ticular points of difficulty in a text, either because a word is difficult or because a syntactic construction is difficult. Detecting these regions and analyzing …

Content truncated to 500 words